For Full-Text PDF, please login, if you are a member of IEICE,|
or go to Pay Per View on menu list, if you are a nonmember of IEICE.
A Dynamic Node Decaying Method for Pruning Artificial Neural Networks
Md. SHAHJAHAN Kazuyuki MURASE
IEICE TRANSACTIONS on Information and Systems
Publication Date: 2003/04/01
Print ISSN: 0916-8532
Type of Manuscript: PAPER
Category: Biocybernetics, Neurocomputing
ANNs, pruning, decaying, competition, classification,
Full Text: PDF>>
This paper presents a dynamic node decaying method (DNDM) for layered artificial neural networks that is suitable for classification problems. Our purpose is not to minimize the total output error but to obtain high generalization ability with minimal structure. Users of the conventional back propagation (BP) learning algorithm can convert their program to the DNDM by simply inserting a few lines. This method is an extension of a previously proposed method to more general classification problems, and its validity is tested with recent standard benchmark problems. In addition, we analyzed the training process and the effects of various parameters. In the method, nodes in a layer compete for survival in an automatic process that uses a criterion. Relatively less important nodes are decayed gradually during BP learning while more important ones play larger roles until the best performance under given conditions is achieved. The criterion evaluates each node by its total influence on progress toward the upper layer, and it is used as the index for dynamic competitive decaying. Two additional criteria are used: Generalization Loss to measure over-fitting and Learning Progress to stop training. Determination of these criteria requires a few human interventions. We have applied this algorithm to several standard benchmark problems such as cancer, diabetes, heart disease, glass, and iris problems. The results show the effectiveness of the method. The classification error and size of the generated networks are comparable to those obtained by other methods that generally require larger modification, or complete rewriting, of the program from the conventional BP algorithm.